The calibration of decision-analytic models. (360G-Wellcome-224853_Z_21_Z)

Background: Mathematical models are often used to inform decisions on how best to spend finite public health resources. These models combine relevant information (in the form of model parameters) about disease processes and relevant interventions to estimate the long-term consequences of the interventions. Unfortunately, some disease processes can not be directly observed but could be estimated through calibration. Model calibration identifies the values of input parameters for which the resulting model outputs are consistent with the observed data. Calibration methods are either Bayesian or non-Bayesian. While Bayesian methods identify the posterior distribution of the parameters, it is easier and quicker to perform non- Bayesian methods. Nonetheless, these methods agree on some occasions. The questions posed by this study are, first, what is the error associated with using non-Bayesian calibration approaches, when are such methods suitable, and what is the value of employing Bayesian methods? Second, what is the error in the net benefits associated with employing imperfect calibration target data, and what is the value of collecting better data? Approach: We will survey relevant literature to identify calibration methods, develop (adopt) methods to answer posed questions, utilise simulations to test developed methods and apply developed methods in a real case study.